Skip to lesson

AnthropicClaude Certified Architect – Professional2026-106.3 Manage stakeholder feedback loops and expectation alignment (including SLAs)

Negotiate SLAs around provider and human limits

Completed by the knowledge check

Learning objective

Negotiate SLAs, SLOs, exclusions, and support ownership that account for provider dependencies, nondeterminism, human handoffs, and fallback behavior.

Core idea

A service commitment describes what stakeholders can rely on and what happens when delivery falls short. It includes more than the model endpoint: retrieval, output checks, human review, fallback, and support all affect the result users receive. Negotiate against that complete service, with costs and responsibilities visible.

Google SRE distinguishes a service level indicator (SLI), the measured quantity; a service level objective (SLO), its target; and a service level agreement (SLA), a commitment with consequences. The measurement design belongs to 1.6.3. Here the architect uses those measures to negotiate coverage, exclusions, consequences, and ownership with product, operations, and the people authorized to accept business commitments.

How it works

Write the terms before agreeing to the target

Create a service agreement schedule that records the following for each commitment:

TermQuestion the agreement must settle
Covered workWhich users, workflow types, operating hours, and input conditions qualify?
MeasurementWhat begins and ends the clock, what counts as success, and what window and denominator apply?
Objective and commitmentWhat internal target triggers intervention, what promise reaches stakeholders, and what consequence follows a miss?
ExclusionsWhich events are outside contractual coverage, who classifies them, and how are they still reported?
Support ownershipWho acknowledges, mitigates, escalates to a provider, authorizes fallback, and updates users?

An internal SLO can be stricter than the stakeholder commitment to leave intervention room. That margin is a negotiated design choice, not a universal ratio. A consequence can be a service credit, a mandatory review, or an agreed remediation obligation; a target without consequences is not automatically an SLA. Product and commercial owners approve the commitment, while engineering demonstrates that its measurement is feasible.

Distinguish exclusions from failure classifications. An unsupported input can be excluded if the boundary is disclosed and observable. A provider outage can still be a failed customer journey even if a negotiated commercial clause excludes it from credits. Report gross user impact and any contractually adjusted measure separately. Otherwise a green adjusted dashboard can conceal an unusable service.

Keep dependencies and variability in the negotiation

Maintain a dependency schedule naming the Claude deployment route, upstream services, available support paths, and contingency owners. NIST GOVERN 6.2 addresses contingencies for high-risk third-party failures. Do not copy a provider’s commitment into your own agreement: different clocks, covered components, or exclusions can make the two promises incomparable.

For the Claude API, the default standard tier provides best-effort availability, and documented rate limits are maximum allowed usage, not guaranteed minimums. Priority Tier capacity commitments are no longer available for purchase; existing commitments continue through their contract end date, and Anthropic directs guaranteed-capacity inquiries to sales. A stronger stakeholder promise needs a funded contingency or capacity negotiated with Anthropic sales; do not quote an availability or capacity guarantee unless the applicable contract states it.

For the Claude API, Anthropic documents HTTP 429 rate_limit_error and HTTP 529 overloaded_error. These are distinct from a business-quality failure in a completed response. The official SDK retries transient failures twice by default; Python max_retries and TypeScript maxRetries configure that behavior. Include actual retry delay in the service deadline and support diagnosis. Raising retries can make the API error rate look better while worsening the user’s completion time.

For nondeterministic outputs, negotiate measured error rates by consequence and task class rather than identical wording or a promise of perfect correctness. Anthropic’s evaluation guidance calls for specific, measurable, achievable, and relevant criteria, often across several dimensions. Record the evaluation population, rubric, review sample, and limitations with the quality commitment. A high aggregate score should not hide an unacceptable failure class.

Define handoff and degraded service explicitly

Give human review its own acceptance and completion clocks. Record the queue owner, staffed hours, acknowledgement deadline, completion target, and what happens when capacity is exhausted. Decide whether the overall workflow clock continues while waiting for the reviewer. If it pauses, make that visible; do not silently remove the slowest part of the experience from the advertised turnaround.

For every fallback, state the trigger, authorized path, quality checks, user notice, remaining deadline, and owner who can disable it. Distinguish an accepted degraded outcome from an ordinary successful outcome. A queued manual case might satisfy a published fallback promise while missing the automated-response objective.

Do not sell the Claude API’s beta fallbacks parameter as outage failover. With anthropic-beta: server-side-fallback-2026-07-01, it addresses safety-classifier declines; rate limits, overload, and server errors return as-is. The parameter is unavailable on the Message Batches API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, and Claude Haiku 5.5 has no server-side fallback. Availability recovery therefore needs an application-owned contingency. See 2.1.2 for model-routing design; the agreement must describe the route the application actually implements.

The support responsibility matrix names one service incident lead even when several teams investigate. Platform operations owns capacity and provider escalation, the domain owner adjudicates quality, and the business operations owner staffs manual work. The service lead remains responsible for stakeholder updates. Provider investigation does not transfer customer communication to the provider.

Key distinctions and trade-offs

Quality versus availability: A completed but unacceptable answer can meet an endpoint availability measure while failing the quality commitment. Negotiate both and disclose their populations.

Recovery versus concealment: Fallback can preserve useful service while increasing latency, cost, and reviewer demand. Count the degraded path and its failures so it cannot silently absorb unlimited work.

Tighter promises versus operating cost: More coverage can require additional staffing, contingency capacity, and testing. Present those costs before approval. An error budget represents allowed SLO misses; agree how its consumption changes release or remediation priorities rather than promising zero misses.

Architecture / decision scenario

A financial-services team wants a Claude-assisted internal case-summary service available around the clock. The sponsor proposes a 99.9% monthly commitment for a completed analyst-approved summary within ten minutes. Analysts are staffed only from 08 to 18 on weekdays, and the budget covers one on-call engineer but no overnight analyst.

The architect presents two feasible options: fund overnight review or separate round-the-clock case intake from staffed-hours approved completion. The service owner selects the second. The illustrative agreement sets a 99.5% monthly intake commitment with a 99.7% internal SLO; approved summaries have a ten-minute p95 target during staffed hours. Overnight intake explicitly promises queue acknowledgement, with review beginning in the next staffed period. Quality is assessed separately on an agreed reviewed sample.

Provider errors count in the gross service report. Bounded retries and a staffed manual queue form the documented contingency; an overnight case is never reported as approved merely because it was queued. The engineer acknowledges critical service incidents within 15 minutes, the operations manager owns analyst capacity, and the service lead updates stakeholders. Commercial owners approve a mandatory service review within two business days after a missed commitment. The original proposal fails because neither provider recovery nor an engineering rota supplies an absent human reviewer.

Common exam traps

  • Committing an availability or capacity figure unsupported by the provider’s best-effort standard tier and maximum, rather than guaranteed, rate limits.
  • Excluding provider failures from every dashboard because they are outside the team’s direct control.
  • Promising end-to-end turnaround while omitting human queue time.
  • Assuming beta fallbacks handles 429 or 529 failures.
  • Naming several support teams without one incident lead and a communication owner.

What to remember

Negotiate coverage, measures, consequences, exclusions, and ownership together. Make dependency failures, variable quality, human queues, and degraded outcomes visible. Promise only what the funded operating model and tested contingencies can deliver.

Knowledge check

References